3D SGG

3D scene graph generation turns sensor data such as RGB-D images, point clouds, and video into graphs whose nodes represent objects or regions and whose edges describe relations between them. Methods combine geometric perception and graph neural networks or transformers with temporal evidence, probabilistic fusion, and language or vision-language models to build hierarchical and open-vocabulary representations. These graphs support robot navigation, object retrieval, manipulation, and scene reconstruction by making spatial structure available for reasoning and queries. Key challenges include incomplete and noisy observations, sparse or inconsistent relation labels, maintaining graphs as scenes change, and balancing detailed representations with real-time computation and reliable evaluation.